4 min readData and AI

AI weather training data: what a model needs from it

An AI weather model can only learn what its training data shows it. Six properties to specify, one public example, and how a custom dataset is scoped, delivered and priced.

Illustration: a translucent three-dimensional model grid hanging over a coastline at night, with one grid cell lit amber around a single raining storm cloud.

First published November 11, 2026 · Updated November 16, 2026

An AI weather model learns the atmosphere from examples. It cannot learn a storm, a sea breeze or a wind ramp that its training data never showed it. So before the architecture, a head of data science has a plainer question: what does the AI weather training data have to contain, and where will it come from? This article sets out six properties to specify, one public example of the scale a project can reach, and how TempoQuest scopes, delivers and prices a custom dataset.

What an AI weather model needs from its training data

Resolution

Grid spacing limits the detail in the training data. If the cells are too coarse to resolve a storm, the model sees only its average over a grid square. If the target is wind across a ridge, cloud over a solar plant or rain on a city catchment, the data has to resolve terrain, coastlines and storms at that scale.

Variables

Train on what you will predict and on what drives it. Wind power needs wind at turbine height, not only at 10 m. Solar needs cloud. Fields taken from one physical model run also agree with each other: wind, temperature and pressure in a grid cell obey the same physics at the same instant. Fields stitched together from separate sources may not.

History

The record has to cover every season and enough of the rare events you care about. Extremes are scarce by definition, so they are the first thing a short record lacks. Hold back whole periods for evaluation. A random split leaks information, because consecutive time steps look almost alike.

Cadence

The time step of the data sets the fastest change a model can learn. A single thunderstorm cell can grow and decay in about an hour, so hourly snapshots miss most of its life. Data every 10 or 15 minutes keeps it.

Consistency

Ask for one model version, one configuration and one grid across the whole record. An archive of operational forecasts changes each time the forecasting system is upgraded, and a learning model can mistake those jumps for weather.

Metadata

Units, map projection, vertical levels, time conventions (UTC, valid time, the accumulation period for rain), the model configuration and missing-value flags. Without them a pipeline guesses, and a wrong guess fails silently.

A public example: Weather 1K

On April 23, 2026, MITRE and The Weather Company announced a collaboration on Weather 1K, a high-resolution AI training dataset that The Weather Company has licensed for GRAF research and development. As the announcement describes it, the dataset has:

  • 1 km spatial resolution over the continental United States;
  • 10-minute increments;
  • vertical coverage from sea level to 62,000 ft;
  • a size of roughly 7 PB.

The announcement names TempoQuest's AceCAST among the modeling tools used to develop Weather 1K. AceCAST runs WRF, the Weather Research and Forecasting model maintained by NCAR, on NVIDIA GPUs.

Weather 1K is here as a public example of the resolution, cadence and volume one project specified. TempoQuest custom data is not a resale of Weather 1K. It is scoped per customer.

How a custom dataset is scoped and delivered

TempoQuest scopes, generates, packages and delivers project-specific high-resolution model outputs and derived datasets for AI training, risk analytics, planning and operations. The scope is where the six properties above become a specification. It covers:

  • Region: the area the data has to cover.
  • Variables: the fields your model predicts and the fields it learns from.
  • Resolution: typical requests are around 1 km or 3 km.
  • Cadence: typical requests are 15-minute, 30-minute, hourly or daily.
  • History: the past period to generate, for a historical dataset.
  • Forecast horizon: how far ahead each run looks, when the data is forecasts.
  • Format and delivery: NetCDF, GRIB, Zarr, CSV, JSON or an API.

Once the scope is agreed, TempoQuest generates the model output, packages it and delivers it in the format you chose.

Bring three things to the first conversation: the model or decision the data has to serve, the observations you will check it against, and the system it has to load into. Write the metadata you need into the scope as well. If the data will be scored against stations, lidar or radar, say so at the start. Our guide to forecast verification explains how that scoring works.

What it costs

Custom Weather Data is priced per project, because the price follows the work. Six things drive it:

  • Domain size. A larger area means more grid columns to compute and to store.
  • Resolution. Finer grids require more computing and storage. See how compute costs are compared.
  • History. Ten years of weather is about ten times the computing and storage of one.
  • Cadence. Output every 15 minutes is 60 ÷ 15 = 4 times the volume of hourly output.
  • Validation. How much checking against observations the project calls for.
  • Integration. How much shaping the delivery needs to fit the systems you already run.

The practical consequence: resolution, cadence and history multiply each other. Decide which of the three your model needs most before you ask for all of them at their maximum.

What to do next

Scope your dataset, or read more about custom weather data from TempoQuest.

Your training dataset

Bring your data requirements.

Share the region, archive, fields and rights your model needs. We review every request.